Computational Models of Human Cognition
Models A model is a means of representing the structure or workings of a system or object. e.g., model car e.g., economic model e.g., psychophysics model 2500 loudness = 830.876 x intensity.233 2000 loudness (sones) 1500 1000 500 0 0 10 20 30 40 50 60 70 80 90 100 intensity
Computational Models Models expressed as computer programs (sequence of instructions) or as complex mathematical equations that require simulation. e.g., air flow e.g., Second Life (3d virtual reality world)
Computational Models of Human Cognition Computer simulation of neural and/or cognitive processes that underlie performance on a task
Goals of Computational Modeling in Cog Sci Understand mechanisms of information processing in the brain Explain behavioral, neuropsychological, and neuroscientific data Suggest techniques for remediation of cognitive deficits due to brain injury and developmental disorders Suggest techniques for facilitating learning in normal cognition Construct computer architectures to mimic human-like intelligence
Why Build Models? Forces you to be explicit about hypotheses and assumptions Provides a framework for integrating knowledge from various fields Allows you to observe complex interactions among hypotheses Provides ultimate in controlled experiment Leads to empirical predictions A mechanistic framework will ultimately be required to provide a unified theory of cortex.
Levels of Modeling Single cell ion flow, membrane depolarization, neurotransmitter release, action potentials, neuromodulatory interactions Network neurophysiology and neuroanatomy of cortical regions, cell firing patterns, inhibitory interactions, mechanisms of learning Functional operation and interaction of cortical areas, transformation of representations Computational input-output behavior, mathematical characterization of computation
Overview Computational modeling Artificial neural networks Modeling performance after brain damage
Key Features of Cortical Computation Neurons are slow (10 3 10 2 propagation time) Large number of neurons (10 10 10 11 ) No central controller (CPU) Neurons receive input from a large number of other neurons (10 4 fan-in and fanout of cortical pyramidal cells) Communication via excitation and inhibition Statistical decision making (neurons that single-handedly turn on/off other neurons are rare) Learning involves modifying coupling strengths (the tendency of one cell to excite/inhibit another) Neural hardware is dedicated to particular tasks (vs. conventional computer memory) Information is conveyed by mean firing rate of neuron, a.k.a. activation
Modeling Individual Neurons x 1 x 2 x 3 x 4. x n w 2 w 1 w 3 w 4 flow of activation w n b y input activities and weights bias output activity
Modeling Individual Neurons Activation function net = i w i x i + b f net f net f net y = f( net) linear binary threshold sigmoid x 1 x 2 x 3 x 4. x n w 2 w 1 w 3 w 4 w n b y input activities and weights bias output activity
Computation With a Binary Threshold Unit Or gate x1 x2 y 0 0 0 0 1 1 1 0 1 1 1 1 y 0 1 1 x 1 x 2
Computation With a Binary Threshold Unit And gate x1 x2 y 0 0 0 0 1 0 1 0 0 1 1 1 y 1 1 1 x 1 x 2
Computation With a Binary Threshold Unit Exclusive or gate x1 x2 y 0 0 0 0 1 1 1 0 1 1 1 0 y = z 1 and not z 2 1 0 1 z 2 = x 1 and x 2 1 0 z 1 = x 1 or x 2 1 1 1 1 x 1 x 2
Feedforward Architectures flow of activity Activation flows in one direction; no closed loops Performs association from input pattern to output pattern big, hairy, stinky run away small, round, orange eat big, round, soft eat small, orange, hairy run away stinky, yellow eat Learning: adjust connections to achieve input-output mapping
Recurrent Architectures Achieves best interpretation of partial or noisy patterns, e.g., MAR M LLOW State space dynamics Attractor dynamics neuron 2 activity neuron 1 activity Learning: establishes new attractors and shifts attractor basin boundaries
Necker Cube Example Each vertex has two possible interpretations. in back in front in front in back Interpretation of one vertex depends on interpretation of other vertices. Constraint satisfaction problem (suitable for attractor net)
Necker Cube Demo See http://www.cs.cf.ac.uk/dave/java/boltzman/necker.html
Supervised Learning in Neural Networks 1. Assume a set of training examples, {x i, d i } e.g., MAR M LLOW MARSHMALLOW e.g., big, hairy, stinky run away 2. Define a measure of network performance, e.g., i E = d i ""y i 2 E 3. Make small incremental changes to weights to decrease error (gradient descent), i.e., w w ji E w ji For multilayered sigmoidal neural networks, gradient descent update has a simple local form (depends on activity of neuron i and error associated with neuron j)
Modeling Neuropsychological Phenomena Michael C. Mozer Mark Sitton Department of Computer Science and Institute of Cognitive Science University of Colorado, Boulder Martha Farah Department of Psychology University of Pennsylvania
Optic Aphasia Neuropathology: unilateral left posterior lesions Deficit in naming visually presented objects, in the absence of visual agnosia and general anomia Nonverbal indications of recognition: sorting, gesturing Naming possible given verbal definition, tactile stimulation, object sounds
Modeling Naming and Gesturing name gesture semantic visual auditory Each arrow represents a processing pathway (neural net) Pathway act as associative memories
Simple Lesion Cannot Explain Optic Aphasia name gesture name gesture semantic semantic visual auditory visual auditory
More Complex Architectures Are Unparsimonious name semantic visual semantics name verbal semantics visual auditory visual auditory name gesture name left hemi semantics right hemi semantics semantic visual auditory visual auditory
Alternative Explanation Partial damage to two systems (Farah, 1990) superadditive effect of damage name gesture semantic visual auditory Simulation by Sitton, Mozer, and Farah (2000)
Neural Network Implementation of Pathway pathway output clean up: recurrent attractor network mapping: multilayer feedforward network pathway input input space mapping clean up output space
Model Dynamics attractor units a(t) s(t) e(t) state units external input attractor unit update equation: â j ( t) = exp( s( t) µ j ) a j ( t) = â j ( t) â i ( t) state unit update equation: 2 β j s i ( t) = h d i ( t)e i ( t) + ( 1 d i ( t) ) a j ( t 1)µ ji j µ j β j : center of attractor j : width of attractor j d(t) t d i ( t) = 1 e i ( t 1) e i ( t) e i ( t) = α e i ( t) + ( 1 α)e i ( t 1)
Key Properties of Neural Network Pathway Gradual convergence of pathway output on best interpretation over time Continuous availability of information from other pathways
Simulation Methodology Define neural activity patterns in visual, auditory, semantic, name, and gesture spaces Pair patterns randomly Train the four pathways to produce correct associations Lesion model Remove 30% of connections in V S and S N pathways Evaluate lesioned model performance
Error Rates by Task task error rate damaged pathways A G 0.0% A N 0.5% S N V G 8.7% V S V N 36.8% V S, S N name visual semantic gesture auditory A N: clean up compensates for S N pathway damage V G: clean up compensates for V S pathway damage V N: effects of damage to V S and S N pathways interact noisy input + internal damage to S N pathway undamaged Interaction would not occur if (a) pathways operated sequentially, or (b) pathways showed no hysteresis quality of output damaged time
Error Rates Based on Relative Damage 100 damage to the S N pathway 70% 60% 50% 40% 30% 20% 10% 50 0 100 50 0 100 50 0 100 50 0 100 50 0 100 50 0 100 50 0 10% 20% 30% 40% 50% 60% 70% damage to the V S pathway
Distribution of Errors for Visual Object Naming 1 0.9 actual chance Proportion of total errors 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 semantic visual vis+sem perseverative other Error type
Interactivity in Brain Damage Neuropsychological disorders have traditionally been explained by a focal lesion to a single processing pathway. Farah (1990) argued that certain highlyselective deficits might have a parsimonious account in terms of multiple lesions with interactive effects. The model illustrates the viability of this account.
Value of the Model Past accounts have claimed the cognitive architecture is complex and unparsimonious. multiple semantics systems or multiple functional pathways to naming Instead, model can explain optic aphasia via a simple cognitive architecture and multiple lesions (each with a single dimension of selectivity). Model can explain other aspects of phenomenon e.g., naming errors tend to be semantic or perseverative, not visual Model might be useful for understanding severity of lesions.